Active learning for human action retrieval using query pool selection

Content-Based Video Retrieval (CBVR) is gaining considerable research interest, inspired by the need to manage the large amounts of video media accumulating on the Internet. In this paper, we verify that the current state-of-the-art retrieval algorithms for CBVR can be improved with active learning. Active learning algorithms query a user for relevance feedback on specific items within the search database, using the additional labeled datapoints to improve the accuracy of the user's original query. We propose a simple CBVR system with SVM relevance feedback, and integrate it with active learning using a simple query pool selection algorithm, based on two co-testing learners. Our experiments demonstrate that such a system performs significantly better with active learning than without, surpassing the state-of-the-art.